A Graph Neural Network Based Conversational Recommendation For Traditional Chinese Medicine Prescription

The combination of Traditional Chinese Medicine(TCM)and artificial intelligence research has gradually emerged in recent years.Some studies focus on the application of intelligent assistance in TCM diagnosis and treatment,aiming to mine patterns and experience from TCM prescriptions and apply them to auxiliary diagnostic tasks such as herb recommendation.The core of TCM diagnosis and treatment lies in "syndrome differentiation and treatment",which requires doctors to grasp the patients’ symptoms through inquiry,identify the internal syndromes by external symptoms,and select appropriate TCM treatments based on the syndromes.However,due to the complexity and ambiguity of TCM syndrome identification methods,the syndrome records of TCM clinical prescriptions are often missing or overly homogenized.Therefore,the TCM herb recommendation method needs to fully exploit the complex relational information within TCM prescriptions and not rely solely on syndromes as the only diagnostic criterion.Furthermore,the existing TCM herb recommendation methods usually require inputting all the patients’ symptoms at once,which demands doctors to have sufficient experience in patient inquiry.Some inexperienced doctors may fail to accurately and completely inquire about the patients ’ symptoms,leading to misdiagnosis when using auxiliary diagnostic applications.This is mainly because existing TCM herb recommendation methods fail to cover the symptom acquisition process.To solve these two problems,the study proposes a graph neural network based conversational recommendation method for TCM prescription.This method can directionally ask patients about their symptoms in multiple turns.Following the TCM inquiry process of "symptom discovery,syndrome summary and herb recommendation",the study constructs a graph structure data based on the complex relations within TCM prescriptions and recommends TCM herbs based on graph neural network.To improve the performance of TCM prescription conversational recommendation,the study proposes a TCM herb recommendation method based on graph neural network for node pre-training to assist the TCM prescription conversational recommendation method.In the TCM herb recommendation node pre-training method,the study constructs a symptom-symptom graph and herb-herb graph based on TCM prescriptions,and aggregates the relational information between symptoms and herbs based on graph convolution networks(GCN).Given a series of symptoms of the patient,the study uses the Transformer module to fuse the embeddings of symptoms and generate the implicit syndrome representation to recommend TCM herbs.In the TCM prescription conversational recommendation method,the study constructs a symptom-herb graph based on TCM prescriptions.Given the initial symptom of the patient,the model uses the GCN module and the Transformer module to dynamically generate implicit syndrome representation for the patient in the current turn.As a state vector,the embedding of syndrome represents the summary of the patient’s condition during the current turn of conversation.In addition,the model selects the most suitable candidate symptoms to ask and the most suitable candidate herbs to recommend in the current turn based on the symptoms confirmed by the patient.Finally,the deep dueling Q-network determines the final action for the current turn from two actions: asking for symptoms and recommending TCM herbs,and waits for users’ feedback.The experimental results on two real TCM prescription datasets demonstrate that the symptom inquiry and herb recommendation method designed based on the TCM diagnostic process in the study have good diagnostic and therapeutic effects.